Viewpoint on preliminary analysis of artificial intelligence-based thyroid nodule evaluation in a non-subspecialist endocrinology setting in thyroidology

Abstract

The recent study regarding the implementation of artificial intelligence-based decision support systems (AI-DSS) in non-subspecialist endocrinology settings provides a timely and essential contribution to the evolving landscape of digital for thyroidologists 1 . By shifting the focus from high-expertise tertiary centers to the point-of-care ultrasound (POCUS), environment of general endocrinology, the authors highlight a critical translational gap: the performance of deep learning algorithms in low-prevalence real-world populations. As such, while the study’s conclusion that non-adjunct artificial intelligence (AI) use may lead to risk overestimation is sobering, several methodological and conceptual nuances warrant further discourse. A primary concern lies in the retrospective nature of the AI analysis conducted on static images captured by non-subspecialists. Furthermore, ultrasound is inherently a dynamic, operator-dependent modality; the “best” diagnostic frame is often selected based on the clinician’s subjective impression. If the initial image acquisition by a general endo-crinologist lacked the most suspicious features of a nodule, even the most sophisticated AI-DSS would be fundamentally limited by the input data. Moreover, the study notes that the AI system detected microcalcifications significantly more frequently than clinicians (10.7 vs. 1.3%), yet it missed comet-tail artifacts—a hallmark of benignity—in almost all cases where clinicians identified them, which suggests that the algorithm may be overly sensitized to high-frequency echoes without the contextual “gestalt” required to differentiate malignant micro-calcifications from benign artifacts. Additionally, the observed low agreement (Kappa=0.026) between the general endocrinologists